TL;DR
This paper proposes a shift from traditional, sequential programming models for computational linguistics towards hardware-based, concurrent processing using Programmable Logic Arrays (PLAs). The authors argue that this "wired-in" approach offers greater descriptive adequacy for understanding language processing as it mirrors the brain's parallel, spontaneous operations, moving beyond mere input-output simulations.
What Problem Does This Paper Address?
The paper tackles a fundamental challenge in computational linguistics: the inadequacy of current models to capture the how of human language processing, not just the what. The authors argue that traditional computational linguistics, relying on sequential, centrally controlled processors (like von Neumann machines), fails to represent the spontaneous, concurrent nature of human cognition. This is analogous to Searle's critique of AI, where a system can perform a task without truly understanding its underlying mechanism. The paper posits that our brains don't run programs on symbol arrays but are "wired" for language, suggesting that building specialized, concurrent hardware architectures is crucial for achieving a more descriptively adequate model of linguistic processes like parsing.
Methods
The core methodological proposal is to leverage Programmable Logic Arrays (PLAs) and related VLSI design techniques to construct specialized hardware for linguistic tasks. The approach involves:
- Neurologically-inspired Net-Linguistics: Building theoretical frameworks based on insights from neural networks.
- Systematic Architecture: Devising idealized architectures by selecting operational primitives (logical operators, PLA constituents like AND/OR planes) and defining basic network modules.
- Concurrent Processing: Implementing linguistic processes (like CF grammar parsing) as complex networks of these modules, where operations occur in parallel rather than sequentially.
- Signal Processing over Symbol Manipulation: Shifting from manipulating abstract symbols to processing and routing signals within the hardware network.
- PLA Implementation: Utilizing PLAs as a key technology for realizing these concurrent networks, allowing for flexible implementation of combinational logic functions that can be "programmed" by configuring the PLA's structure.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Hardware as a Model: The most significant insight is the proposal to use hardware construction (specifically PLA nets) as a more adequate model for language processing than software simulation on universal machines. This directly addresses the "how" of linguistic competence.
- Concurrency is Key: The paper strongly emphasizes that human language processing is inherently concurrent and spontaneous, a stark contrast to the sequential nature of traditional computer architectures.
- Descriptive Adequacy: The authors argue that achieving "descriptive adequacy" requires not only explaining what humans do (e.g., parse sentences) but also how they do it, which necessitates modeling the underlying concurrent processes.
- PLA as a Building Block: PLAs are presented as a practical and flexible technology for implementing these concurrent networks, allowing for significant changes to functions without major redesign.
- Signal Flow over Symbol Manipulation: The shift from symbolic manipulation to signal flow within the network is highlighted as a more accurate representation of how neural systems operate.
Key Results
The paper presents illustrative examples of how concurrent signal flow can be realized in PLA networks for linguistic tasks, specifically:
- Grammar Realization: Demonstrates how a CF grammar can be translated into a PLA network where units correspond to grammar rules and symbols, and signal flow represents parsing.
- Concurrent Parsing Example: Figure 4 and the accompanying table illustrate the concurrent signal flow for parsing the string "abd" using a specific grammar, showing how different parts of the parse are processed in parallel.
- Bottom-Up Processor Exploration: The authors indicate plans to investigate realizations of concurrent signal flows for bottom-up processors, though a general method is still under development.
Limitations and Future Work
- General Method for Bottom-Up Processors: The authors acknowledge that a general method for specifying bottom-up processors in terms of networks is still needed.
- Complexity of Networks: While illustrating simple examples, the paper implies that the complexity of real-world linguistic processing will require more extensive research into defining and managing these concurrent networks.
- Foundational Concepts: The paper expresses hope that several years of experience with VLSI designs for such systems will lead to fruitful foundational concepts for process-oriented linguistics.
- Further Research: The authors explicitly state that much more research is necessary to fully realize the potential of this approach.
Summary updated